Claude can help HR teams draft job descriptions, summarize engagement survey themes, prepare interview guides, and turn messy notes into clear documentation. Those are good uses. The uses that need guardrails are the ones where model output could influence a decision about a person — who gets interviewed, how someone's performance is characterized, how an employee-relations case is handled. This article lays out recommended practice for an HR addendum to your layered AI policy. It is not legal advice: automated or AI-assisted employment decisions are regulated differently across jurisdictions, so review the final addendum with employment counsel.
The decision-influence line
The core rule most HR addenda converge on: Claude may assist with documents and summaries; it may not screen, rank, score, or recommend outcomes for individual people. Concretely:
Recruiting. Drafting a job posting or structuring interview questions is low-risk. Feeding a stack of résumés in and asking which candidates to advance is a different category entirely — it puts model behavior directly in the hiring decision, where errors and skewed patterns are hardest to detect and most consequential. If your organization ever considers that use, it belongs in your highest risk tier with formal bias evaluation and counsel review before any pilot.
Performance support. Helping a manager turn their own observations into clearer written feedback is assistance. Asking the model to characterize an employee's performance from raw data is delegation of judgment — the manager must remain the author of the assessment, not the editor of the model's.
Employee relations. Case summaries and timeline reconstruction can save hours, but investigation conclusions, credibility judgments, and disciplinary recommendations should remain entirely human. Many teams also restrict ER case material to a dedicated, access-controlled deployment or exclude it from AI tools altogether.
Data guardrails
Employee data — compensation, health-related accommodations, disciplinary history, immigration status — usually sits in your organization's most restricted classification tiers. The addendum should map those tiers to explicit allowed/prohibited answers per approved deployment, using your data sensitivity matrix so nobody has to guess.
Platform facts worth knowing when you calibrate: for Anthropic's commercial products, inputs and outputs are not used for model training by default — the exception is explicit opt-in, such as submitting feedback. That exception matters operationally: feedback submissions may be retained for up to 5 years, so an HR addendum should instruct staff never to use feedback/report buttons on content containing employee data. For the Claude API, the stated default is deletion of inputs and outputs within 30 days; stricter zero-data-retention arrangements exist by contract. On Amazon Bedrock and Google Cloud, the cloud provider is the data processor — confirm equivalent controls with your provider.
Oversight that fits HR
Three lightweight mechanisms cover most of what an HR addendum needs beyond the rules above. First, a named reviewer: any AI-assisted document that will enter an employee's record gets read and owned by a human before filing. Second, disclosure norms: decide when candidates or employees should be told AI assisted a process, and default to transparency where the answer is unclear. Third, periodic sampling: HR leadership reviews a sample of AI-assisted outputs each quarter for tone, accuracy, and consistency across groups — a small, regular check that catches drift long before an audit would.
Where to go next
Read evaluating AI outputs for bias before any use case touches candidate or employee comparisons, and designing human oversight for the review-gate patterns. For the underlying data-handling facts, see the sources below.